feat: Add gradient testing for Flash Attention 2

[crStiv]

This commit adds gradient testing for Flash Attention 2 (FA2) implementation.
The test ensures that gradients computed with FA2 match those computed with
the eager implementation within specified tolerance thresholds.

Key changes:

• Added gradient testing in train mode
• Compare gradients between eager and SDPA implementations
• Use same tolerance thresholds as forward pass comparison
• Properly cleanup gradients and restore eval mode

This addresses the TODO comment in test_modeling_common.py and improves
test coverage for FA2 implementation.

[ArthurZucker]

Nice! I think the test can be simplified : no longer needs 2 attention models as you can just set the config._attn_implementation for newer models!